i have been following Chinese models for about two years now because they are open-weight and qwen is fun to run on my kubernetes cluster, the news about this Apache licensed model complete with a recipe to make it again with potentially different ingredients is making we want to abandon them for something more ideologically sound and way more interesting
this new model, couldn’t universities rebuild it with different contexts and study it in ways you can’t reproduce in other models? like, is k2 horizons a good scientific foundation on which to study machine learning?
or am i just lacking way too much context and falling for hype?
This seems to score well on benchmarks. Better than any previous “open blueprint release”. While an exact ordered blueprint on static data provides reproducibility, it’s not obvious that a different ordering, or more recent versions of webpages/github repos would produce inferior results. I don’t really see a huge area of study available from the blueprint, though other labs may find areas of improvement/inspiration in their approaches.
K2 wasn’t the first.
China had CPM-1 in 2020 that was a 2.6B model.
GPT-Neo (Mar 2021) EleutherAI replicates early GPT
PanGu-α (Apr 2021) Huawei’s 200B model
WuDao 2.0 (May 2021) BAAI’s massive 1.75T sparse MoE model
GPT-J (Jun 2021) EleutherAI release 6B
Meta released OPT in 2022, BLOOM was in 2022, GLM 130B by Tsinghua/Zhipu was in 2022 etc
K2 horizon is very recent, nowhere near the first. Different labs have different computational innovations worth studying. DeepSeek is crazy efficient and doing very novel things. Mistral in France is making very compact local friendly models that are fun and easy to fine tune and merge on consumer hardware.
My question would be if it uses stuff like CommonCrawl as training material. Given their size I assume they used anything incl. unethical training data, but at least admit it?
The only models I ever found that even tried to only resort to ethically obtained data, being FOSS etc. were the tiny ones from PleIAs. And as expected they’re completely useless.
So far I concluded that an “ideologically sound” LLM is impossible due to lack of training data. Unless your ideology allows to steal stuff.
I’m assuming a lot of its training data is synthetic and distilled from Chinese models that were themselves trained from pirated data and distilled from American models trained on pirated data. It would be quite remarkable if the training data involved no piracy whatsoever. Then again, it’s open source, so I suppose it would be essentially reversing a reverse Robin Hood.
Someone ping me once the IFM releases the training data for K2 Horizons and I’ll add it to the list. Currently it says “training data will be released to the public” and I can’t find any big downloads, so it’s not open-source yet.
But sure, big achievement. We only got a handful somewhat recent, usable in size/performance, true open-source LLMs.
I try most models, that get support in llama.cpp or forks. IIRC I wasn’t overly impressed with this one. Slower than it should be, mostly because its KV cache is huge.
I keep falling back to qwen. Though bonsai did give it a good run for a while.
But I only have an old 8gb card.
I have a bit more memory, so I jump between the two qwen 3.8s.
Glimmer muse was a good contender in the MoE space for a few days. Might be more viable on 8gb.
Qwen is what I’ve found to work the best for me so far. I haven’t done an exhaustive search or anything, but the mixture of experts 35b models seem to work pretty well.
I’ve got a little more vRAM to play with, 20GB. It still struggles with giving it enough context to be useful for agentic stuff.
